基于双通路融合网络的用户视频情感识别OA

Facing challenges from dynamic content,diverse themes,and sparse emotion distributions in user-generated videos,this paper developed a dual-path fusion network(DPFNet).The core of DPFNet was a dual-path encoding structure with attention mechanisms.The main path modeled temporal dependencies of visual embeddings,while an auxiliary path introduced emotional auxiliary text to reconstruct and constrain visual features for adaptation to diverse themes.Then a dual-path decoder processed both paths in parallel to capture emotion-relevant visual content.Finally,a dynamic gating mechanism adaptively fused the decoded features to suppress task-irrelevant interference.Experiments on two public datasets show that DPFNet achieves classification accuracies of 60.2％and 63.7％,respectively.DPFNet demonstrates more efficient recognition using only visual modality input while maintaining comparable classification performance,significantly reducing model complexity and computational costs.